[CIG-MC] Citcom runtime on Lonestar vs Ranger (Teragrid): Insights/experience?

[Eh Tan]

Walter Landry wrote:
> Finally, memory pressure may have an effect.  The AMD's generally have
> better memory performance, but there is more contention since there
> are 16 cores accessing the same memory, rather than 4 cores accessing
> 4 separate banks of memory.
>   

The memory pressure (bandwidth limit) is indeed a major bottleneck for 
CitcomCU and CitcomS. Generally, FEM codes require high memory bandwidth 
to solve the sparse matrix inversion. Longstar machine has two dual-core 
CPUs, and Ranger machine has four quad-core CPUs. All cores in the CPU 
share the same memory bandwidth. As a result, Ranger has effectively 
half memory bandwidth per core, compared to Longstar. This can affect 
the speed by 5-10%.

I've done a speed benchmark showing the effect of memory bandwidth on 
our cluster, which has two quad-core CPUs.
http://www.geodynamics.org/cig/Members/tan2/scaling/document_view


Cheers,
Eh Tan